Parametric memory scoring evaluates an AI model's static training data by capturing native brand recall, using token-level log-probabilities and exploring alternative token paths to quantify unprompted association confidence.

In contrast, grounded memory scoring forces real-time web retrieval and URL context tools to establish verified, source-backed facts about a brand, enabling reconciliation via the Fact & Gap Checker to detect hallucinations, confirm accurate recall, and isolate brand knowledge gaps.